Literature Title: Deep Computer Models of Cognition
Paper Title and Abstract
Building Blocks of Humanesque Cognition
Within the space of knowledge representations in the Artificial Intelligence literature, the existence of some subset is chiefly justified for their amenability to automation, others are well automated but also have implications to the modeling of cognition, while a still smaller subset is directly motivated by cognitive problems. This paper gives a review of those knowledge representations and ideas most relevant to the computational modeling of human cognitive problems, which we term humanesque cognition; it is an attempt to re-present and synthesize the literature’s diverse corpora of ideas, and an effort to reinvigorate these ideas by re-articulating their relevance to problems of humanesque cognition, including, remembrance, instinct, rationality, attention, understanding, learning, feeling, creativity, consciousness, and homeostasis.
Building Blocks of Humanesque Cognition “Do not stay in the field!
Nor climb out of sight.
The best view of the world
Is from a medium height”
- Friedrich Nietzsche, The Gay Science
An appeal for a medium height. Early in the history of Artificial Intelligence, it was thought that singular knowledge representations— be they logic-based symbolic systems, or artificial neural networks – would quickly vanquish the human intelligence problem, but more than fifty years later, this dream has not (some would add, yet) materialized. After some reflection on the nature of knowledge representations, and with the benefit of hindsight, we observe that the very act of formalizing a knowledge representational idea into a computable entity requires some compromises to its expressive power; these limitations have been variously noted as “ontological commitments” (Davis, Shrobe & Szolovits, 1993), and degradative transformations into “formal syntactic manipulations” (Searle, 1980). We understand that it is this class of limitations which prevents a knowledge representation from being good at modeling all cognitive problems. So to bring this point to its natural conclusion, there is a sense that perhaps the human intelligence problem might best be tackled piecewise, bringing to bear multiple representations (Minsky, 1986, forthcoming), pairing each particular cognitive problem with the knowledge representation best suited to it. Some purists would argue that brains work on a common representational substrate, however, our work is not to replicate a brain, it is, more pragmatically, to draw inspiration from human capabilities to build humanesque cognition. To clarify, we want to follow a path not of deep AI, or weak AI, but of medium AI. It is scruffy because the field collectively does not yet know enough to do things neatly.
Motivated by this spirit of a medium height, we regret that so many ideas and knowledge representations in the Artificial Intelligence literature that were originally conceived with grand cognitive problems in mind, have slowly evolved into vanilla computational tools now belonging only to weak AI. From the initial seeds of this paper, we hope to begin a fresh dialogue to reinvigorate the space of AI knowledge representations by re-articulating their relevance to problems of humanesque cognition, including, remembrance, instinct, rationality, attention, understanding, learning, feeling, creativity, consciousness, and homeostasis. We cannot claim that this paper is completely unique in its motivation or in its structuring, or that the arguments contained within are completely rigorous, but there is a sense that this kind of gestalt narration about AI is perhaps more lacking than it ought to be. If you will, think of this an attempt to poeticize rather than to didacticize.
On that note, we dive straightaway into the following problems of humanesque cognition: remembrance, instinct, rationality, attention, understanding, learning, feeling, creativity, consciousness, and homeostasis.
The word ‘remember’ itself derives from the etymon, memor in Late Latin, meaning mind; so quite literally it means to put back again into the mind. Yet when we examine the work on case-based reasoning (CBR) (Schank & Abelson, 1977; Riesbeck & Schank, 1989; Leake, 1996) and memory-based reasoning (MBR) (Stanfill & Waltz, 1986) in the literature, we regret that the sanitized vanilla versions of these computational structures which have been embraced by weak AI and pop AI only pay lip service to the spirit of remembrance; more specifically, the recollection of a case or a memory in MBR entails only the rote selection of a symbolic expression from a base of symbolic expressions, rather than the reactivation of some contextual pattern-state of some computed mind. In defense of computing, we know that a mind is a hard thing to instantiate computationally, particularly because there are so many unknowns about how such a program might be written or how this mind would even behave; however, one solution is simply to build some simpler mind for the particular purpose of exploring the memory question, not a human mind, but simply some computational apparatus which is capable of remembering with any kind of contextual and attentional sophistication. Luckily, much of the work outside of the vanilla versions of CBR and MBR are full of clues and enticements toward a more cognition-centric computational modeling of remembrance.
Properties of memory. Cognitive scientists who have studied the human memory faculties like to make the useful bifurcation of ‘memory’ into reflexive memory (RM), and long-term episodic memory (LTEM). While long-term episodic memory deals in salient, one-time events and must generally be consciously recalled, reflexive memory is full of automatic, instant, almost instinctive associations. In this section on remembrance, we address chiefly LTEM, and defer discussion of RM to the next section on instinct.
One of the most dominant ideas with respect to LTEM is Tulving’s Encoding Specificity Hypothesis (1983), which stated simply, portrays a LTEM as including within its boundaries a set of contextual features such as affect, scents, sights, sounds, persons in the scene, et cetera. The intersection of these contextual features generate a specific encoding signature for that episode. Later, that memory is possibly retrievable by any of the encoded features, including the contextual ones, although the presence of more features will improve the chances of recalling that particular episode; thus we think of LTEM as content-addressable.
Case-based reasoning and rote remembrance. In case-based reasoning (Schank & Abelson, 1977; Riesbeck & Schank, 1989; Leake, 1996), a case is a computational unit typically organized as a frame, with slots and slot values. Sometimes, a case frame simply has two types of slots allowing it to behave as an expert-systems rule: antecedents, and consequents (and this particular variety of limited CBR is called knowledge-based reasoning). The act of remembrance within this framework is initiated by the formulation of a search frame describing the present situation, and a matching algorithm, often heuristic and “fuzzy,” identifies some subset of all cases in the case-base as being “relevant” to the current situation. This simple formulation of remembrance has been successfully applied to a broad range of problem domains varying from medical diagnosis (where memories are expert-crafted rather than augmented experientially by the computer program) to modeling creative processes, e.g. (Turner, 1994), in which past situations might suggest relevant solutions to a present situation.
One of the limitations to the CBR approach to remembrance is in fact a limitation to all symbolic systems, and that is the problem of ontological commitment (Davis, Shrobe & Szolovits, 1993). In a cognitively interesting memory system, each memory might be encoded drawing from a larger repertoire of slots than those typically available to CBR systems, including plenty of contextual features. Rather than having a flat organization, the slots themselves would have many inter-relationships and be partially structured by inner hierarchies. The space of cases or memories, while flat and homogeneous in typical CBR, would have more interesting topologies in a cognitively interesting memory system, perhaps distorted by many idiosyncrasies and biases. And the memory itself would evolve, with cases or memories merging, and others garbage collected. After all, in a humanesque cognition, minds undergo forgetting, belief revision, and theory change.
Contextually fluid remembrance. Recognizing that people remember in much more sophisticated fashion than symbolic approaches like CBR, there are some proposals in the literature which aim to increase the contextual sophistication of representing memories. In The Society of Mind (1986), Minsky proposes a more connectionist approach to memory aimed at tackling the context problem by improving the number of interconnections between symbolic features involved in memories. In Minsky’s theory of memory, K-Lines are conceptualized as threads which weave together many symbolic features, and may also activate other k-lines. Anytime a symbolic feature is triggered, all the k-lines threading through it are also potentially triggered. Minsky also proposes two specializations of K-Lines called nemes and nomes. The neme idea is particularly relevant to the present discussion. We have thus far described a need for more sophisticated remembrance—one which pattern-activates the mind, inducing it into a particular state; unfortunately rote memory recall a la CBR is a far cry from achieving this. However, Society of Mind theory’s neme idea is closer to the mark. In Examining the Society of Mind (2003), Singh describes two useful nemes, polynemes, and micronemes.
For example, recognizing an apple arouses an ‘apple-polyneme' that invokes certain properties within the color, shape, taste, and other agencies to mentally manufacture the experience of an apple, as well as brings to mind other less sensory aspects such as the cost of an apple, places where apples can be found, the kind of situations in which one might eat an apple, and so forth.
Micronemes provide ‘global' contextual signals to agencies all across the brain. Often, they describe aspects of our situation that are too subtle for words and for which we otherwise have no more determinate concepts, such as specific smells, colors, shapes, or feelings.
– Singh, Examining the Society of Mind (2003)
Thusly viewed, polynemes enable pattern-activation type of remembrance, and micronemes draw subtle contextual factors into remembrance, many of which exist liminally or fall below the threshold of articulation.
Minsky also offers mechanisms for varying the strengths of relevance of particular details to a memory. Level bands (Minsky, 1986, p. 86), for example, allow symbolic features sitting along the same K-Line to activate with different strengths, depending on which level-band of detail that particular feature sits in. For example, given the remembrance of a kite, features like “paper” and “string” sit in the relevant level band of detail and get activated more strongly than features in either the “upper fringe” band (e.g. “toy,” “gift”) or “lower fringe” band (e.g. “diamond shape”). Activations in the fringe bands are weak and may be overwritten, but they are useful as default assumptions. Level-band theory affords a reconciliation of representational granularity, and there is a sense that such reconciliations are key to building contextually fluid memory systems.
Blurring and forgetting. One of the features of CBR limiting its cognitive plausibility is that cases are largely static and homogenous entities; each experience is afforded exactly one case, and there is no account of how these memories evolve, combine, or degrade. In The Muse in the Machine (1994), Gelernter raises this very issue and his FGP (Fetch-Generalize-Project) computer program is an attempt to address blurring and forgetting. In FGP, a present situation prompts the recall of some set of memories. These memories are compacted into a memory sandwich and a set of features present in all the memories emerge in an overlay. This overlay is essentially a memory blur, and one interpretation of Gelernter’s theory is that memories can evolve when many memories combine into an overlay, and the overlay is kept but the original memories are garbage-collected. Over the course of a memory’s lifetime, aspects of the memory which do not emerge in any overlays may be deemed to fall below some forgetting threshold, and thus be garbage-collected to economize the memory.
Narrative memory. So far we have considered memories as static moments, however, some memories are animated, including narrative memory— a remembrance of a dynamic sequence of events. In Scripts, Plans, Goals, and Understanding (1977), Schank and Abelson put forth a script as a knowledge representation for a narrative memory. A script is usually a simple linear causal sequence of events, which taken together encapsulates a simple narrative. Many interesting elaborations to this basic unit of narrative memory have been explored. For example, in Daydreaming in Humans and Computers (1990), Mueller explores imaginative memories in which the space of all narrative memories, both what has happened and what might conceivably have happened or could happen, is explored. We see the reconciliation of actual narratives and potential narratives as a positive exploration of narrative memory’s fluidity.
The word instinct is a provocative one, and those philosophically minded will no doubt demand clarification. Instinct is a suitcase concept which has been taken to refer to different things, including inter alia, evolutionary dispositions, genetic dispositions, habitualizations, intuitions, and hard-wired behaviors. From these different meanings taken together, three essentials seem to emerge: 1) an instinct is a consequence or side-effect of some mechanism or machinery such as a mind; 2) there is a sense that its representation is somewhat muddled, indirect, or emergent (ruling out programming instincts by explicit symbolic rules, which somehow seems to violate the spirit of the idea); 3) an instinct manifests as a reaction without deliberative thought.
Instinct and reactive AI. In the literature, the line of research on reactive AI, including insect AI seems to centrally address the notion of instinct. In “Intelligence Without Representation,” (1991a) and “Intelligence Without Reason,” (1991b), Brooks lays out what is a manifesto for a low-representational approach to creating intelligent robots using hidden markov models (HMMs) organized into subsumption hierarchies to govern behavior. Intelligent creatures built this way might be called instinctive because the three essentials apply here: 1) a creature’s behavior is a consequence or side-effect (since non-symbolic HMMs do not demonstrate intention) of the HMM machinery; 2) the representation of behavior is not explicit, but rather, arises out of some emergent coordination between distributed mechanisms; 3) a creature’s behavior is reactive, and involves no deliberation.
The success of reactive AI in robotic creatures is a statement that not all intelligence must be explicitly programmed, or can be reasoned through protocols of rationality. Some intelligent behaviors inevitably arise out of idiosyncrasies in a representation, without an explicit symbol or rule which could be pointed to as holding responsibility.
Instinct as reflexive memories. Instinct is not only relevant to creatures, but also to minds. Idiosyncratic tendencies of a mind such as “personality” might be considered instinctive, and so might “intuition” and “habit” (if one accepts that what results from ontogeny and acculturation can be considered instinct still).
Here we wish to pick up on a representational idea we first raised in the previous section on Remembrance. Earlier, we distinguished between long-term episodic memories, and reflexive memories. While long-term episodic memory deals in salient, one-time events and must generally be consciously recalled, reflexive memory is full of automatic, instant, almost instinctive associations. Tulving equates LTEM with “remembering” and reflexive memory with “knowing” (Tulving, 1983).
We raise the issue of “reflexive memories” here because it exhibits some of the nice properties of instinct, and provides a potential account of how some instinct could be acquired or honed through classicalconditioning. First, a reflexive memory is an immediate reaction which requires no deliberation (if could have once required deliberation but through the course of habituation, become cached as a reflex). Second, the reflexive association learned between stimulus and reaction depends on the idiosyncratic features of the training set responsible for the learned response. Third, the explanation behind why the particular reflexive association was learned is certainly unavailable, outside of pointing back to all the different exposures in the conditioning process which reinforced that association; in other words, the association emerges out of experience with the training set.
Instinct out of forgotten deliberations. Of course, reflexive memories could arise in an alternate fashion. Consider that given some stimulus, a response was formulated after some deliberation. However, over repetitions of the stimulus and the selection of the same response, the stimulus-response pair became habituated, and in fact the original deliberation responsible for formulating the original response is no longer valid, but the stimulus-response pair persists out of habit. In this account, reflexive memories are formed when original deliberations are forgotten, but the reflex continues out of habit.
But why instinct? Of all the “interesting cognitive problems” presented in this paper, the topic of instinct most warrants justification. At first, it may seem to be a lesser cognitive problem more relevant to lesser animals than to human cognition. However, there are actually some high-level very human capabilities which seem to be at least in part instinctive, such as personality, and aesthetic tastes. How is it, for example, that a person can know quite instantaneously whether or not they like a cuisine or a type of music, yet not be able to articulate the reasons why? Many insights might be gained but looking analogically at how behaviors emerge out of the system dynamics of reactive AI systems, and at how reflexive memories are formed through classical conditioning or from once deliberative processes.
Rationality is perhaps the most well-covered cognitive topic in the literature, and in fact, the logical tradition factors prominently into AI’s heritage. Concepts ranging from logical reasoning, to goals, to planning, to action selection in autonomous agents epitomize rational thinking. Some however stress that rational thinking is only one type of thinking. Gelernter, for example, describes a spectrum of thought (1994) where rational thinking like deduction lives at the high focus end, dreaming and discursive thought at the low focus end, and creative, analogical thinking as somewhere in the middle. Ultimately, rationality represents an idealization of human thought, inspired by logic, economic theory, and decision-theoretics. Because this ideal is widely understood and obeyed socially, rational thinking is useful because it becomes a protocol for communication, socially-defensible behavior, and prediction of the behaviors of others.
The appeal of logic. Part of the appeal of rational thinking is the general perception that it is the discovery of some objective truth, and for this reason, logic is often regarded as quintessential framework for objective truth. For its part, logic offers various assurances about the condition of truth. The most common type of logic, the first-order predicate calculus (FOPC), has an ever appealing property called monotonicity, which assures that the truth of any proposition will not change when new axioms are added or old axioms deleted. Traditional logics are based on deduction, which is a method of inference which is quite exact and foolproof.
While logics are good at modeling formal discourses, they are arguably ill-equipped for real world common sense reasoning, which calls for more heuristic reasoning in the face of incomplete, inexact, and potentially inconsistent information. Nevertheless, devout logicians have attempted to model the common sense reasoning problem via the FOPC, beginning with McCarthy’s suggestion (1958), and culminating in the Cyc Project (Lenat, 1995) in which common sense knowledge is represented in two million logical assertions.
Economic theory and decision-theoretics. Much of what is considered the protocol of “rational thinking” is thinking within a decision-theoretic framework, with the guidance of economic theory (Good, 1971). Generally, a rational agent should make decisions which maximizes goodness. For example, the action selection process in autonomous agents research is modeled according to economic laws (Maes, 1994). Given the current state of the world, and a set of goals, the ideal next action for an agent should result from a cost-benefit-risk analysis, where goals are scored by their priorities, or organized into goal-subgoal hierarchies. Other classic examples of idealized rational thought in the AI literature include the SOAR cognitive architecture’s model of goal selection and impasse fixing (Newell, 1990; Lehman et al., 1996); and the planning systems for the STRIPS blocks world solver (Fikes & Nilsson, 1971) and the early mobile robot Shakey (Nilsson, 1984). In the AI planning literature, utility theory from economics is responsible for the filtering principle, i.e. constrain reasoning by eliminating lines of reasoning which conflict with goals, and the overloading principle, i.e. try to construct a plan which satisfies the greatest number of goals (Pollack, 1992).
Rationality and mindreading. One of the most compelling uses of the rationality principle is for predicting or explaining the behaviors of other social actors, often referred to as mindreading in the Cognitive Science literature (but not to be confused with the psychic powers of fortune-tellers and the like). The way in which humans understand and predict the behaviors of other social actors is best captured by Dennett in his postulation of the Intentional Stance (1987). Dennett postulates that in order to predict the behavior of a social actor, one would view that social actor as an intentional system possessing beliefs and desires, whose behaviors result from rational deliberations on those beliefs and desires based on primarily on the economic principle of utility. For example, I believe I have 1.00$ to spend and I believe I see an ice cream, and I desire an ice cream; therefore the intentional stance would predict that I intend to buy the ice cream. Mindreading and the associated concept of theory of mind also emerges in the autonomous agents literature, because autonomous agents often find it necessary to predict the actions or explain the beliefs or desires of other agents, and a popular representation for this kind of agent modeling is the Belief-Desire-Intention model of agency (Georgeff et al., 1998), closely tied to Dennett’s formulation.